Setting up an AI assistant at work: a 6-step guide

Generative AI tools are a click away, and it is tempting to paste in a contract, a price list or a customer email to save time. A business that wants to benefit from AI without losing control of its information would do well to set up its own assistant: connected to its documents, reserved for its own staff, with clear permissions.

Here is how such an assistant works, followed by the six steps to set it up.

How a business AI assistant works

Four building blocks work together:

• The language model (the AI’s “brain”) understands the questions and writes the answers. It has a lot of general knowledge, but knows nothing about your business.
• Searching your documents fills that gap. Before answering, the assistant looks for the relevant passages in your indexed documents, then uses them to write its answer. This is called RAG (retrieval-augmented generation).
• Connectors give access to your sources: Google Drive or OneDrive, SharePoint, network folders, email, CRM, ERP, databases…
• Actions let the assistant do things: draft and send an email after confirmation, update a spreadsheet, create a record in a CRM.

Step 1: choose two or three concrete use cases

Start small, with frequent tasks that are easy to check: answering internal questions from your procedures, finding a clause in contracts, drafting a reply to a customer, summarising a long document. An assistant that does one specific job every day is far more likely to be adopted than one that is supposed to “do everything”.

Step 2: identify and tidy up your sources

The assistant will never be more reliable than the documents it consults. List the sources your use cases need, then sort through them: an outdated procedure or three versions of the same price list will lead to wrong answers. This clean-up directly improves the quality of the answers.

Step 3: decide who can access what

The golden rule: the assistant must never give anyone information they could not otherwise see. In practice:

• one account per user, and role-based permissions (the accounts team does not necessarily see HR files);
• if possible, sign-in with the company’s existing accounts (SSO), so people do not end up with yet more passwords;
• a log of conversations, so you know who asked what.

Step 4: secure the data

A few checks before you connect anything:

• The AI model provider: read its terms to find out where your data is processed and to check that it is not used to train its models.
• The GDPR: if the assistant processes personal data (customers, employees), the regulation applies just as it does to any other tool. Limit the data to what is useful, inform the people concerned and put contracts in place with your providers. The CNIL (the French data protection authority) publishes recommendations dedicated to artificial intelligence on its website.
• The EU AI Act: since 2 February 2025, its Article 4 has required businesses that use AI systems to make sure their staff have a sufficient level of AI literacy. A short training session on how to use the assistant, including its limits, is a step in that direction.
• The most sensitive data: the assistant can be installed in your own environment (on-premise) rather than on a hosted platform.

Step 5: test with a small group

Open the assistant to a few users for a few weeks. Note the questions it answers badly: most of the time, a document is missing or a source is ambiguous. Fix that, then roll it out more widely.

Step 6: plan the full budget

An AI assistant has an upfront cost and a monthly cost:

• Upfront: the core package (the assistant, accounts, permissions, document indexing, deployment), then the connectors, actions and modules you need.
• Every month: the platform that hosts the assistant, user licences, maintenance (updates, fixes, compatibility with new models and connected tools) and, depending on usage, AI consumption beyond what is included.

For reference, at Ibsia the core package of a business AI assistant starts at €6,900 excl. VAT, connectors at €800 excl. VAT, modules at €900 excl. VAT and maintenance at €99 excl. VAT per month; the platform and licences are quoted on request.

The limits to be aware of

• AI can be confidently wrong. Searching your documents reduces this risk but does not remove it: keep a human check on important decisions, and a confirmation before any action (as is the case for sending emails).
• Quality depends on the sources. One up-to-date document is worth more than ten drafts.
• Cost follows usage. The more the assistant is used, the more it consumes; keep an eye on consumption for the first few months.

Setting up an AI assistant: summary

1. Two or three concrete use cases to start with.
2. Sources that are sorted and up to date.
3. Per-user permissions, and a log of conversations.
4. Protected data: provider, GDPR, AI Act.
5. A test with a small group.
6. A budget that includes the monthly costs.

Want to talk it through? Take a look at our business AI assistant, or start with task automation if your need is more focused.